docs(agents): the MuJoCo GL pin is cited by the name the test defines

[cagataycali]

What

Two citations of test_no_module_scope_windowed_gl_default - the Pinned by line in AGENTS.md and a :func: role in tests/test_examples_mujoco_gl.py's own module docstring - now spell the name the file defines, test_no_module_scope_platform_bound_gl_default. A new grader, tests/test_agents_md_pinned_by_citations_resolve.py, holds every path::Name coordinate in AGENTS.md to that rule.

Why

The test was renamed when its rule was widened from the windowed pair (cgl, glfw) to every platform-bound backend, and the two citations were not. The Pinned by coordinate is what a reader copies into pytest to run the pin, so as it stood the rule read as enforced while the pin it named could not be run:

$ pytest tests/test_examples_mujoco_gl.py::test_no_module_scope_windowed_gl_default
ERROR: not found: .../tests/test_examples_mujoco_gl.py::test_no_module_scope_windowed_gl_default

Measured over AGENTS.md at 0b93bc2a: 16 path::Name coordinates, 15 resolving and this one not.

The bare :func: role in the test module is out of scope for tests/test_docstring_xref_roles_resolve.py by that grader's own design (a bare name has no decidable target in general - measured here, 566 of 4199 bare roles across the graded trees name a sibling class's method or a cross-module class, so widening it would be noise). A path::Name pair is decidable, because the path names the file to read, which is why the pin lives in a separate grader rather than in that one.

How the pin works

• The graded set is derived from AGENTS.md itself (regex over <path>.py::<Name>), so a coordinate added later is graded without editing the test.
• Resolution reads the cited file's AST rather than importing it, so a pin whose module needs an optional extra is still graded.
• A dotted Class.method resolves segment by segment through direct bodies; a single segment is accepted at any depth, because AGENTS.md also cites a nested function (simulation.py::_job) as a reading coordinate.
• Reports the line, the coordinate, and which half failed (missing file / no such definition / file does not parse).

Verification

• New grader: 18 passed. Mutation: reverting the AGENTS.md fix fails test_every_citation_resolves naming line 2288 and the stale name; planting a nonexistent file and a nonexistent name fails naming each.
• Whole-tree graders (scripts/check_whole_tree_graders.py roster, 145 modules) run on this branch and on main in the same environment: 61 failed / 4310 passed / 151 errors on both, identical failure sets - all absent-extra (lerobot, mujoco), none attributable to this change.
• ruff check, ruff format --check, mypy clean on the touched files.
• Intake: check_merge_base_overlap.py --paths reports no open pull request editing a behaviour-bearing path in this set (fix(simulation/isaac): Isaac Sim backend #3343 shares AGENTS.md as prose only, +1 line elsewhere in the file).

Review rounds

• Round 0: opened as draft; fragment pushed before marking ready (AGENTS.md PR Workflow step 3).

[yinsong1986]

Summary

Fixes a stale test citation in two places - the Pinned by line in AGENTS.md and the :func: role in tests/test_examples_mujoco_gl.py's module docstring - both of which named test_no_module_scope_windowed_gl_default after the test had been renamed test_no_module_scope_platform_bound_gl_default, and adds tests/test_agents_md_pinned_by_citations_resolve.py to hold every path::Name coordinate in AGENTS.md to resolvability. Verified against the head commit: the renamed test exists at tests/test_examples_mujoco_gl.py:270, no other citation of the old name survives outside provenance notes, the new grader passes on the head tree (18 passed), and its regex sweep collects all 16 path::Name coordinates in AGENTS.md with every one resolving - including the nested-function (simulation.py::_job) and dotted-class cases the resolver claims to handle.

What's good

• The graded set is derived from AGENTS.md itself, so a coordinate added later is graded without editing the test (the derived-roster convention from PR Workflow step 2).
• Resolution reads the cited file's AST rather than importing it, so pins whose modules need optional extras are still graded, and test_the_graded_set_is_not_empty guards against the regex silently matching nothing.
• Test names describe behaviour rather than provenance, new files are ASCII-only, and the changelog fragment landed in the same push as the diff (step 3's timing rule).